23 gateways and open-model hosters hand out real credits at signup — and only 8 can prove zero data retention

Asked (summary):

Find niche, newly-funded AI infrastructure startups: a highly filtered list of LLM gateways (routing/proxy layers) and open-source model hosters that explicitly grant free API credits or builder grants on signup or application. Log zero data retention (Yes / No / Available on request) and link directly to pricing, docs, or applications.

The qualified set covers 23 providers and programs verified against first-party pricing, billing, startup-program and security pages as of the 2026-09-17 research cutoff: 12 grant credits automatically or on a recurring schedule, 11 require an application or claim form. Only concrete gratis balances count — dollar credits, tokens, GPU hours or requests — never a plain free request-rate tier. Evidence spans 22 independent credit-source hosts, so the set is not single-sourced.

The credit landscape: dollar size by access method, coloured by ZDR status

Access

ZDR

ZDR yes — official policy documents a no-payload-storage mode (default or toggle) Available on request / configuration No — payloads retained under strict definition Not publicly verified

Bars use a log dollar scale from $0.10 to $50,000. Awards in tokens, hypercredits or GPU capacity are listed in the lower band — no invented dollar conversions. Hover a bar for eligibility and retention detail.

The five largest dollar awards are FriendliAI up to $50,000, Together AI up to $50,000, Baseten up to $25,000 plus $2,500 for Model APIs, and Novita and Inference.net up to $10,000 each — all application-gated. friendli.ai together.ai baseten.co

RunPod's startup program is the largest non-dollar award — up to 1,000 free H100 GPU hours and 1,000,000 serverless requests — yet it is the only "No" on strict ZDR: sync results persist 1 minute, async 30 minutes, job data until TTL. runpod.io docs.runpod.io

Ten providers hand credit without any application — from Beam's $30/month and Merge's $10 card-free first month down to Hugging Face's $0.10/month — plus Charm Hyper's 100 hypercredits/month and Inception Labs' 100M tokens. beam.cloud merge.dev

Default ZDR without any toggle exists only at Baseten (sync inference), Fireworks (open-model services), Hugging Face's routing layer and Modal endpoints; Together, Nebius, Requesty and Merge need organization-level switches — and Fireworks' Response API still stores 30 days unless store=false. docs.fireworks.ai modal.com

Large selective grants

FriendliAI (up to $50K, sized to your current inference bill), Together AI accelerator ($15K/$30K/$50K tiers), Baseten ($25K + $2.5K), Novita and Inference.net ($10K), GMI Cloud SCALE ($500 then up to $4,500), Requesty ($1,000 for YC/20VC/Tapestry/TinyVC startups), Modal (tiered, undisclosed amounts) and RunPod (capacity). All demand an application and most require venture backing or early stage proof.

Automatic smoke-test balances

Beam $30/month, Merge $10 first month without a card, Cerebras $5 after payment verification, Clarifai $5 with phone verification, Vercel $5/month, Fireworks $1, SiliconFlow $1, Hugging Face $0.10/month for free users, Charm Hyper 100 hypercredits/month, Bytez recurring credits (amount undisclosed), Inception Labs 100M tokens, Scaleway 1M tokens plus a €100 voucher.

Retention-control caveats

"Yes" here means a documented no-payload-storage mode — not a "we don't train on your data" promise. Nebius and Together ZDR are org-level toggles, off by default. Merge counts only when org ZDR routing is on and gateway payload logging is separately disabled. Requesty needs per-key logging off or Enterprise org ZDR. Fireworks' Response API defaults to 30-day storage. Baseten async inputs are held until processing completes.

Strict exclusions

Hyperbolic was cut: its referral credit requires the referred user to top up $5. Tavily was cut as a search API, not a gateway or hoster. NVIDIA was cut because an official forum answer says build.nvidia.com removed its credit system. Plain free rate tiers and playground-only access never qualified.

The complete qualified matrix

Provider & type	Free credit / program	ZDR status	Offer link	Notes

Method: 23 qualified providers/programs assembled from first-party pricing, billing, startup-program, grant and security pages as of the 2026-09-17 research cutoff; 24-row candidate matrix minus Hyperbolic (paid top-up required), Tavily (search API, off-category) and NVIDIA (credit system removed per official forum), plus Merge Gateway and SiliconFlow verified on their own pricing pages; 28 product-level retention rows checked for the ZDR column. Dollar bars show the stated maximum award on a log scale; token, hypercredit and GPU-capacity awards are kept separate without dollar conversion. ZDR "Yes" requires an official documented no-payload-storage mode; "Not publicly verified" means no public strict policy was found, not an assertion either way. Evidence spans 22 independent credit-source hosts.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.